Hybrid Deep Learning Model for Retinal Disease Detection with XAI

This study introduces a new way to automatically spot retinal diseases by analyzing Optical Coherence Tomography (OCT) images. Model accuracy was evaluated using the OCT2017 and OCT-Duke-Srinivasan-2014 datasets after training and testing on the OCT-C8 dataset. These datasets include a mix of retinal conditions like AMD, CNV, CSR, DME, DR, DRUSEN, MH and healthy retinal images. Before training the data in the model, some preprocessing steps were applied, like resizing images, normalizing them, augmenting the data to add variety, and reducing noise to help the model learn better. The proposed model is a hybrid deep learning system that combines Convolutional Neural Network (ResNet50) with a Recurrent Neural Network (RNN) based Long-short-term memory (LSTM). The ResNet50 part is used to extract important features from the OCT images, and the LSTM part helps to understand the spatial and sequential patterns. To train the model, the Cross-Entropy Loss function was used. After extracting features using ResNet50, the LSTM network processes these features for the final classification. The model was also tested on the OCT2017 and OCT-Duke-Srinivasan-2014 datasets to see how well it performs on new data, and it gave better results than other existing models. The model achieved a sensitivity of 97.54%, a specificity of 98.96%, and an AUC of 0.9976%. The model also applied an explainable AI method called Gradient-weighted Class Activation Mapping (Grad-CAM), which helped us to understand which parts of the image the model was focusing on while making decisions. This kind of explainability is important if anyone wants to use the system in real hospitals. Although the study faced some issues like small retinal details and noisy images, this method shows good promise for building a reliable and automatic retinal disease diagnosis system that can help doctors and make eye checkups easier and faster.